Translation tools such as Google Translate can cause misunderstandings
The ability to translate non-native pages open in Chrome is useful, but there are some parts that can only be determined if the translation is accurate. Dr. Vinay Prabhu and colleagues at Carnegie Mellon University point out that there are cases where women become men as a result of translation, and 'forbid violence' becomes the opposite, such as 'order violence'. doing.
Google translate bias investigations --YouTube
Dr. Prabhu et al. Announced 'Did they direct the violence or admonish it? A cautionary tale on contronomy, androcentrism and back-translation foibles. Precautions regarding reverse translation) ”. The treatise.
(1 / n) Our paper 'Did they direct the violence or admonish it? A cautionary tale on contronomy, androcentrism and back-translation foibles' will be presented #AfricaNLP workshop today ???? @eaclmeeting @GradioML Notebook: https: / /t.co/0aytQSwBOh— Vinay Prabhu (@vinayprabhu) April 19, 2021
???? ️ https://t.co/k9G5Jlfelw pic.twitter.com/ABBoKT18uY
The first thing they focused on was the 'she survival rate,' which indicates a masculistic bias. This is a numerical value that indicates whether the word 'she' remains when 'reverse translation' is performed to retranslate the translated text in the original language.
(2 / n) In this initial work, we focus on two facets of shortcomings of #GoogleTranslate— Vinay Prabhu (@vinayprabhu) April 19, 2021
The first, is the androcentrist default-to-male bias measured via 'She survival rate' of languages: The odds that the pronoun'She' will survive a backtranslation journey to a language pic.twitter.com/PiEpdsFtLk
When I went to the African language, there was a big difference that the 'her survival rate' was extremely low in the Niger-Congo language family and quite high in the Afro-Asian language family. This is thought to be rooted in 'whether or not there is a gender difference in language'.
(3 / n) In the African context, we observed a Niger-Congo v / s Afro-asiatic schism develop that has roots in the way the languages are gender-loaded (Cue Guiora Scale / Word atlas of languages chapters 30-32) pic.twitter.com/bTsE5yVjwk— Vinay Prabhu (@vinayprabhu) April 19, 2021
The next thing they focused on was the mistranslation of antonyms. The survey was conducted longitudinally in all 109 languages.
(4 / n) The second facet we explore is the downstream effects of mistranslating contronyms (Janus word / enantiosemy). Firstly, we did a longitudinal study across all the 109 languages— Vinay Prabhu (@vinayprabhu) April 19, 2021
Idea credits: @ abidlabs https://t.co/2fXoo7o0MD pic.twitter.com/KwIy6quhyt
In addition, in order to clarify the weak points, he said that he also conducted a cross-sectional survey of more antonyms in a single language.
(5 / N) And Then, Proceeded To Also Perform Latitudinal Exploration Across Many Contronymous Phrases Across A Single Language To Uncover More Frailties. Pic.Twitter.Com/6793oV7dcx— Vinay Prabhu (@vinayprabhu) April 19, 2021
(6 / n) This paper took me through many a beautiful by-lane of psycholinguistics (The Shapir-whorfian linguistic relativity debates!) And introduced me to the works of Dr. Rashid Jahan (Credits: @iawoolford ) and Yoruba linguists such as @aremuadeolajr pic.twitter.com/iQWZQE6Fhn— Vinay Prabhu (@vinayprabhu) April 19, 2021
The full version of the treatise will be published soon.